Conference abstracts

Session B4 - Learning Theory

July 15, 15:00 ~ 15:25 - Room B6

Learning in Nature

Nisheeth Vishnoi

EPFL, Switzerland   -   nisheeth.vishnoi@epfl.ch

In this talk, I will present two algorithmic connections between nature and humans that we recently discovered. The first is an equivalence of the dynamics of the slime mold and a popular algorithm in machine learning -- the iteratively reweighted least squares (IRLS) method. The second is between a distributed learning dynamics observed among social animals (including humans) and the multiplicative weights algorithm.

Not only have similar learning algorithms been independently chosen by human intelligence and nature (via evolution) to solve problems in entirely different context, these connections lead us to the first proof of correctness of the damped IRLS method and a novel distributed and low-memory implementation of the classic MWU method.

Joint work with L. Elisa Celis (EPFL), Peter Krafft (MIT) and Damian Straszak (EPFL).

View abstract PDF



FoCM 2017, based on a nodethirtythree design.